An integrated model combined conventional radiomics and deep learning features to predict early recurrence of hepatocellular carcinoma eligible for curative ablation: a multicenter cohort study

doi:10.21203/rs.3.rs-5226011/v1

Download PDF

Research Article

An integrated model combined conventional radiomics and deep learning features to predict early recurrence of hepatocellular carcinoma eligible for curative ablation: a multicenter cohort study

https://doi.org/10.21203/rs.3.rs-5226011/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Aim This study aimed to develop and validate a model (DLRR) that incorporates deep learning radiomics and traditional radiomics features to predict ER following curative ablation for HCC.

Backround Hepatocellular carcinoma (HCC) is the most common primary liver malignancy. Ablation therapy is one of the first-line treatments for early HCC. Accurately predicting early recurrence (ER) is crucial for making precise treatment plans and improving prognosis.

Methods We retrospectively analysed the data of 288 eligible patients from three hospitals—one primary cohort (centre 1, n=222) and two external test cohorts (centre 2, n=32 and centre 3, n=34)—from April 2008 to March 2022. 3D ResNet-18 and PyRadiomics were applied to extract features from contrast-enhanced computed tomography (CECT) images. The three-step (ICC-LASSO-RFE) method was used for feature selection, and six machine learning methods were used to construct models. Performance was compared via the area under the receiver operating characteristic curve (AUC), net reclassification improvement (NRI) and integrated discrimination improvement (IDI) indices. Calibration and clinical applicability were assessed via calibration curves and decision curve analysis (DCA), respectively. Kaplan-Meier (K-M) curves were generated to stratify patients based on progression-free survival (PFS) and overall survival (OS).

Results The DLRR model had the best performance, with AUCs of 0.981, 0.910 and 0.851 in the training, internal validation, and external validation sets, respectively. NRI and IDI tests indicated that the DLRR model outperformed the DLR model (AUCs of 0.910 and 0.874; P < 0.05) and the Rad model (AUCs of 0.910 and 0.772; P < 0.05). Although the AUC of DLRR was slightly lower than that of the combined model (incorporating DLRR and clinico-radiological features), there was no significant difference (AUCs of 0.910 and 0.914; P > 0.05). Additionally, the calibration curve and DCA curve revealed that the DLRR model had good calibration ability and clinical applicability. The K-M curve indicated that the DLRR model provided risk stratification for progression-free survival (PFS) and overall survival (OS) in HCC patients.

Conclusion The DLRR model noninvasively and efficiently predicts ER after curative ablation in HCC patients, which helps to categorize the risk in patients to formulate precise diagnosis and treatment plans and management strategies for patients and to improve the prognosis.

HCC

Ablation

Eearly recurrence

Radiomics

Deep learning

Machine learning

Primary liver cancer is the second leading cause of cancer-related death globally, and its incidence is increasing. The number of new patients with primary liver cancer worldwide is expected to exceed one million by 2025^[1,2]. Hepatocellular carcinoma (HCC) accounts for approximately 90% of primary liver cancers and poses a significant challenge to global health because of its poor treatment outcomes^[3]. Ablation has been shown to be a minimally invasive and first-line treatment for early-stage HCC, achieving therapeutic outcomes similar to those of surgical resection, but the recurrence rate within five years is approximately 70%, which is higher than that of surgical resection, limiting its applicability^[4–7]. Early recurrence (ER) leads to increased mortality and a poor survival prognosis than late recurrence^[8,9]. Since HCC patients are mostly diagnosed by imaging without histopathological examination and lack risk factors related to recurrence, such as Microvascular Invasion(MVI)^[9], predicting ER after ablation has been a difficult clinical research task. Therefore, there is an urgent need for a reliable technique that can noninvasively and efficiently predict ER after ablation in HCC patients to formulate a precise diagnostic plan for patients as well as a management strategy to improve prognosis^[10,11].

Radiomics was initially proposed in 2012 as a comprehensive method for analysing medical images to quantify imaging phenotypes and advance precision medicine^[12]. In recent years, predictive models based on radiomics have become increasingly popular for assessing the risk of HCC recurrence after surgical resection^[13–16]. We constructed an integrated model combining radiomic features with clinico-radiological features, which demonstrated effective predictive performance for early recurrence after liver resection for HCC in patients with cirrhosis^[17]. However, there are few predictive models for ER after ablation therapy for HCC. Beleu et al analysed CT texture features in the ablation area and identified five features as independent predictors of local recurrence risk; the model had a C-index of 0.73, which is suitable for assessing the risk of local recurrence in HCC patients after radiofrequency ablation therapy^[18].

Deep learning has great application prospects in the field of medical image analysis; however, it faces the risk of insufficient data and overfitting. Deep learning networks can extract features on their own, and the combination of such features with traditional radiomic methods can effectively improve the accuracy and robustness of the corresponding models. Li et al. named this method deep learning-based radiomics (DLR) and tested and verified its accuracy in prediction. Compared with traditional radiomics, the DLR method yielded better results for almost all relevant indicators^[18]. Convolutional neural networks (CNNs) are the most widely used deep learning architectures in medical image analysis. Among them, ResNet-18 has shown great potential in the prediction of various diseases. Wang et al. developed and validated a ResNet-18-based multimodal model based on preoperative MR and CT images to predict MVI in HCC; the model achieved an AUC of 0.819, indicating high prediction efficiency^[20]. Wei et al. developed a hybrid model to predict muscle invasion in patients with bladder cancer via deep learning radiomics and traditional radiomics (DLRR), achieving AUCs of 0.884 in the internal validation cohort and 0.862 in the external validation cohort, demonstrating outstanding predictive accuracy^[21]. To date, no studies have reported the application of DLRR in predicting ER after curative ablation therapy for early HCC.

In this study, we aimed to construct a model based on contrast-enhanced computed tomography (CECT) images via the DLRR method to predict ER in HCC patients following curative ablation to assist in the development of early-stage HCC treatment and prognostic management programs and provide a feasible solution for the application of deep learning to the ablation treatment of early-stage HCC.

Patient selection

This multicentre retrospective cohort study was approved by the Ethics Management Committee of the First Affiliated Hospital of the University of Science and Technology of China (2021-RE-043). We enrolled a total of 288 patients with early-stage HCC who received curative ablation at three centres: the First Affiliated Hospital of the University of Science and Technology of China (Center 1), the Second Hospital of Anhui Medical University (Center 2), and the Taizhou Hospital of Zhejiang Province (Center 3) from April 2008 to March 2022. The inclusion and exclusion criteria for patient selection are shown in Fig. 1A. At Center 1, patients were randomly assigned to a training set and an internal validation set at a 7:3 ratio. Patients from Center 2 and Center 3 were combined to form the external validation set.

Ablation was performed by the same experienced team of the respective hospitals under ultrasound guidance as described in eAppendix 1. If there were multiple lesions, ablation was performed one by one. All patients underwent CECT, and the imaging data were collected via the picture archiving and communication system (PACS) of the respective hospitals. Specific CT equipment information is shown in eAppendix 2. The inclusion criteria were as follows: (1) clinical diagnosis of HCC according to the noninvasive criteria established by the American Association for the Study of Liver Disease based on distinct imaging features^[22]; (2) single tumour diameter ≤ 5 cm, multiple tumours ≤ 3, each diameter ≤ 3 cm; (3) refusal to undergo hepatectomy or liver transplantation; and (4) patient management involving curative ablation only. The exclusion criteria were as follows: (1) tumours invading blood vessels, bile ducts, adjacent organs, distant metastasis, or other malignancies; (2) a prior history of HCC treatment, such as hepatic resection, transarterial chemoembolization (TACE), targeted therapy, or radiotherapy; (3) the absence of CECT imaging data or CECT conducted more than 1 month prior to ablation; and (4) a follow-up duration of less than 2 years. The detailed workflow is illustrated in Fig. 1B.

Clinical data collection

Before initiating the data collection process, all relevant staff members at the participating centres underwent a training session on data extraction. The medical records of eligible patients were meticulously reviewed. Clinical and laboratory data were systematically collected via standardized forms. To ensure relevance and consistency, only the data obtained within one week before ablation were considered. Another reviewer randomly assessed and validated 30% of the collected data. To account for potential variations across participating centres, all laboratory measurements were standardized. Extreme outliers—values significantly higher or lower than the norm—were flagged for review. These outliers were re-evaluated by the hospital's lead researcher or the designated chief physician to confirm their validity and exclude input errors.

Image information collection

CECT images were obtained from the hospital PACS in DICOM format. Two physicians (reader 1 and reader 2) from each centre independently evaluated the CECT images and focused on the following eight semantic features: (1) tumour margin; (2) tumour capsule; (3) intratumoral vessels; (4) tumour growth; (5) intratumoral necrosis; and (6) peritumoral enhancement. When multiple lesions were present in the patient's liver, we evaluated the largest tumour. Some example images and specific imaging semantic features are explained in eAppendix 3.

Image segmentation

Physician A, who had 10 years of experience interpreting abdominal CT scans, used ITK-SNAP (version 3.6.0; www.itksnap.org) software to segment the images. The volume of interest (VOI)—comprising either the entire tumour or the residual liver excluding vessels or bile ducts—was delineated layer by layer on arterial, portal, and delayed-phase images. In cases with multiple lesions, the largest lesion was chosen for segmentation. Throughout the delineation process, Physician A was blinded to the clinical data of the patients.

To extract the histological features of the peritumoral images, the peritumoral VOIs were processed via the Python morphological erosion and expansion algorithm, which automatically constricted the boundary of each lesion inwards by 5 mm and expanded it outwards by 3, 5, and 10 mm, respectively. To ensure the reproducibility of the radiomic features, physician A and physician B, with 10 years of experience in reading abdominal CT images, performed the above procedures again after two weeks. The results of the repeat extraction were used to calculate the interclass correlation coefficient (ICC).

Feature extraction via radiomics and deep learning

Feature Extraction

Python version 3.8.4 (https://pypi.org/project/pyradiomics/) was used to extract radiomic features (eAppendix 4) for three image phases (arterial phase, A; portal phase, P; delayed phase, D) and six regions (tumour, residual liver, 5 mm-eroded and 3 mm-, 5 mm-, and 10 mm-extended peritumoral regions). To quantify the differences between different phases, we calculated the differences in the radiomic features between the A and P phases and between the D and P phases in the six regions (delta-radiomics). For each patient, 1539 imaging features were extracted. After removing features with a variance close to 0, the total number of features was reduced to 1,512 in the tumour and residual liver regions and 1,507 features in the other regions, resulting in a total of 45,290 features per patient. For more details, please refer to Table S1.

The settings for 3D ResNet-18 prior to feature extraction are described in eAppendix 5. In the deep learning feature extraction process, the extraction area was the same as that used in the radiomic process. The number of features extracted from each region was 512, and a total of 9216 deep learning features were extracted for each patient from three phases and six regions (Table S3), ensuring the consistency and comparability of the data.

Radiomic features and deep learning feature screening

To prevent overfitting, we employed a three-step feature selection process for feature screening. First, we selected features with an ICC greater than 0.8 and standardized these features via Z score normalization. Second, we used LASSO regression to select features with nonzero coefficients. Third, if the number of features exceeded one after LASSO selection, we further utilized recursive feature elimination (RFE) with a decision tree (DT) kernel to determine the optimal number of features at each stage. To obtain more representative features, each group of features underwent the above feature selection process. The specific feature selection process and results are shown in Tables S2. Finally, we calculated the Rad-score on the basis of the weighted regression coefficients of the radiomic features derived from LASSO.

In addition, we also conducted the above three-step feature selection processes for features extracted through deep learning. The specific feature selection results are shown in Table S4. We calculated the DL-score on the basis of the weighted regression coefficients of the deep learning features derived from LASSO.

Model building and comparison

We built seven models utilizing different features. The clinico-radiological model (Cli) was constructed from the clinical and imaging features resulting from univariate and multivariate analyses, the Rad-Score calculated from the radiomic features was used to construct a radiomic model (Rad), and the DL-Score calculated from the deep learning features was used to construct a deep learning radiomic model (DLR). To identify potentially better prediction models, we constructed four integrated models based on the clinico-radiological features, Rad-Score and DL-Score, including the clinico-radiological and Rad-Score integrated model (CR), the clinico-radiological and DL-Score integrated model (CDL), the Rad-Score and DL-Score integrated model (DLRR), and a comprehensive model combining all features (Combined).

We constructed seven models via 6 machine learning algorithms, including support vector machine (SVM), logistic regression (LR), random forest (RF), K-nearest neighbour (KNN), light gradient boosting machine (LightGBM) and Xtreme gradient boosting (XGBoost) algorithms. The most appropriate algorithm was selected on the basis of the characteristic data of the different groups to ensure the objectivity of the results. To improve the generalization ability of the models and better evaluate their performance with small sample sizes, we used 5-fold cross-validation for the model hyperparameter selection and model training process. Moreover, the process was performed on the training set only to avoid data leakage. In addition, we excluded models with an AUC greater than 0.1 between the training set and the internal validation set to avoid overfitting and underfitting. The best model with the highest AUC in the internal validation set was selected.

Model validation and clinical application

To demonstrate the calibration ability and clinical applicability of the best model, we generated calibration curves and performed DCA on the training, internal validation and external validation sets.

To evaluate whether the best model can effectively predict progression-free survival (PFS) and overall survival (OS) for risk stratification, we utilized the maximum Youden index from the internal validation set as the optimal cut-off value for prediction outcomes in both the training and validation cohorts. Patients were categorized into low-risk and high-risk groups, and the 2-year PFS and 5-year OS rates were analysed via Kaplan-Meier(K-M) survival curves.

Follow-up

All patients were followed up regularly after discharge. The first follow-up was one month after the ablation procedure, during which the local therapeutic effect was evaluated. The other follow-ups were every three months or six months after ablation. During these follow-ups, the serum AFP level, liver function, and abdominal ultrasound examination were conducted, and CECT, MRI, or ultrasound angiography might be used to monitor for recurrence if necessary. The starting point of this study was defined as the time at which the ablation procedure was performed, and the primary endpoint was ER. ER was defined as the emergence of a new intrahepatic lesion or metastasis within two years postablation, with the lesion displaying imaging features typical of HCC or being confirmed through histopathological analysis. Curative ablation was defined as the absence of tumour necrosis enhancement on dynamic CECT, MRI, or CEUS. The last follow-up date for this study was March 31, 2024.

Statistical analyses

Statistical analysis was performed via R software (version 4.3.0). In the training set, variables with more than 20% missing data were excluded; otherwise, multiple imputation algorithms were employed to handle the missing data. To improve model interpretability, continuous variables were converted into binary variables via threshold values from receiver operating characteristic (ROC) curves. Categorical variables are presented as frequencies and percentages and were analysed via chi-square tests or Fisher's exact tests. Variables showing a p value of less than 0.1 in the univariate analysis were included in the multivariate analysis for further variable selection. The models' predictive performance was assessed through the area under the curve (AUC), accuracy (ACC), positive predictive value (PPV), and negative predictive value (NPV). Model comparisons were made via the net reclassification improvement (NRI) and integrated discrimination improvement (IDI) indices. Calibration curves and decision curve analysis (DCA) were used to evaluate model calibration and clinical utility. Kaplan-Meier (K-M) survival analysis was conducted, and log-rank tests were used for curve comparisons. A two-tailed p value of less than 0.05 was considered statistically significant.

Baseline characteristics of the study cohorts

In Centre 1, a total of 222 patients were included in the final analysis, with 155 patients assigned to the training cohort, and 64 patients were assigned to the internal validation cohort. In Centres 2 and 3, 66 patients were included in the independent external validation cohort. The clinico-radiological characteristics of the patients are shown in Table 1. As of the final follow-up, the ER rate among the HCC patients was 38.2% (110/288).

The clinical information and imaging characteristics of the HCC patients in the training set, the internal validation set and the external validation set were not significantly different (P＞0.05). The results of the univariate and multivariate analyses are presented in Table 2. Age, AFP, cirrhosis and the presence of intratumoral vessels were found to be independent predictors of ER in patients with HCC (P＜0.05).

Results of radiomics and deep learning feature selection

Among the three imaging phases, 20 radiomics features (17 radiomics features and 3 delta-radiomics features) and 27 deep learning features were found to be associated with ER in patients with HCC after curative albation. The features and associated feature coefficients are shown in Table S5.

Results of model construction and comparison

The Cli model was constructed from four clinico-radiological features, while the Rad-Score calculated from 20 radiomic features was used to construct a radiomic model (Rad), and the DL-Score calculated from 27 deep learning features was used to construct a deep learning model (DLR). The ROC curves for each model are shown in Figure 2. According to a comparative analysis of the models, the DLR model outperformed the Cli and Rad models, with AUCs in the training set and internal validation set of 0.908 (0.855-0.961) and 0.874 (0.795-0.953), respectively.

With respect to the integrated models, Table 3 shows the best machine learning method corresponding to each model and their performance metrics. The ROC curves in the training set and internal validation set for each model are shown in Figure 2. Moreover, we determined the NRIs and IDIs for these models and found that the prediction effects of the combined model and the DLRR model were better than those of the other integrated models (Table 4). The AUCs of the two models in the training set were 0.996 and 0.981, and those in the validation set were 0.914 and 0.910. Although the AUCs of the combined model in both the training set and the internal validation set were better than those of the DLRR model, the differences were not significant, indicating that the addition of clinico-radiological features did not improve the predictive performance of the model. Considering that the DLRR model is more convenient than the combined model, as it can make predictions without the need to collect complicated clinical and imaging data, we chose the DLRR model as the optimal model for the following analyses.

Model validation and clinical application

The DLRR model had an AUC of 0.981 in the training set, 0.910 in the internal validation set, and 0.851 in the external validation set (Figure 3). Thus, the DLRR model has good generalizability. The DLRR model had good calibration and overall net benefits in both sets (Figure 4).

Risk stratification

The DLRR model could accurately stratify patients based on PFS and OS (both <0.0001) (Figure 5). Additionally, we demonstrated the process and corresponding results with two examples in Figure 6.

Accurately predicting ER after curative ablation of HCC is crucial for guiding precision therapy and improving patient prognosis^[23].Although the application of machine learning in radiomics has provided new predictive tools, the need to incorporate various clinical and biomarker data hinders its further promotion. Therefore, there is an urgent need to develop a more efficient and convenient model to predict ER in early-stage HCC patients. Our study collected preoperative CECT data and extracted features from three phases and six regions via DLRR methods. We constructed an integrated model based on associated features simultaneously. The results showed that the model had significant advantages in predicting ER, with AUCs of 0.981, 0.910, and 0.851 in the training, internal validation, and external validation sets, respectively. The K‒M curves and corresponding cumulative risk curves clearly stratified both the training and validation sets (P < 0.05).

This study identified clinical and imaging features, and four variables were found to be independently associated with ER according to univariate and multivariate analyses, among which three factors, age, AFP and cirrhosis, were consistent with the results of previous studies^[24–26]. Bosi et al. confirmed that intratumoral vessels are associated with the growth of HCC^[27]. The AUC values in the training and validation sets of these four features were 0.712 and 0.690, respectively.

Radiomics was used to extract features from different phases and regions, and delta-radiomics was obtained by subtracting radiomic features from different phases. A total of 45290 features were extracted from each patient, and seventeen radiomic features and three delta-radiomic features were ultimately obtained after three feature screening steps were applied. Features from the portal and delayed phases accounted for 82% (14/17) of the radiomic features and 74% (20/27) of the deep learning features. This result was similar to that of the study by Yuan et al., who constructed a model for predicting recurrence based on three-phase CT images and reported that the portal and delayed phase features accounted for 75% of the 20 related features^[28]. The reason may be that although arterial phase images clearly reveal the abundant blood supply of early HCC tumours, the radiomic features of the portal and delayed phases could better reflect the microvascular structure and perfusion within the liver; this microlevel information is crucial for evaluating the prognostic effect of HCC ablation therapy.

Radiomics features in the surrounding tumour area play crucial roles in predicting early tumour recurrence^[29]. In this study, 88% (15/17) of the features were in the peritumoral region, and 82% (14/17) of the features were within the 10 mm peritumoral range. The results demonstrated the importance of the peritumoral regional characteristics in predicting recurrence, similar to the findings of Zhou et al. and Shi et al.^[30,31]. Therefore, in the absence of pathological information, CECT images of the 10 mm peritumoral area can reveal features closely associated with early recurrence^[32]. Further analysis of the seventeen radiomic features revealed that 30% (6/20) of the features were related to coarseness, a feature that describes the texture roughness of an image. To explore the distribution of this feature on the corresponding CT images, we used the "feature mapping" method to map the coarseness under different filters (Fig. 7) (eAppendix 6 for specific feature mapping steps). The results show that the coarseness features were mainly distributed in the boundary regions, with a large distribution in the 10 mm areas. Therefore, we suspected that this feature may be correlated with the heterogeneity of the peripheral microenvironment of the tumour and may influence tumour behaviour and patient prognosis. We plan to perform further research in the future to confirm this conjecture.

CNNs are the primary deep learning network for extracting features and can find deeper features than traditional radiomic methods can find, which reflect more important tumour information and make prediction easier^[33–35]. In this study, 9,216 deep learning features were extracted for each patient from multiple phases and regions via 3D ResNet-18. After feature selection, 27 deep learning features were ultimately used to construct the DL model. 89% (24/27) of the features were located in the peritumoral region, and 63% (17/27) of the features were within 10 mm of the area. Further analyses revealed that the AUCs of the predictive model in the training and validation groups were 0.908 and 0.874, respectively, which were significantly better than those of the clinico-radiological model. This finding was consistent with the findings of Wu et al., whose study reported a C-index of 0.695 (0.561–0.789) for predicting ER in the validation cohort, outperforming the clinical model^[36]. The superior performance of the model may be due to the extraction of features from multiple phases and regions, which may contain more prognostic information.

To increase the predictive efficiency of the model, Ma et al. integrated deep learning and radiomics features, as well as integrated intratumoral and peritumoral regions, to provide more valuable information for predicting the therapeutic response of non-small cell lung cancer patients to chemoradiotherapy^[37]. Zhang et al. constructed a model based on 9 deep learning features and 17 radiomics features to diagnose meningioma, achieving an AUC of 0.943 (0.873-1.000) in the test cohort, indicating potential clinical value for assisting doctors in preoperative tumour diagnosis^[38]. There have been no studies on the application of DLRR in predicting the prognosis of patients with HCC after ablation therapy. The AUCs of the DLRR of our model were 0.981 and 0.910 in the training and validation sets, respectively. Through NRI and IDI comparative analyses, we found that the DLRR model was significantly superior to the CR, DL, Rads, and Cli models. However, the combined model, which incorporates clinico-radiological data into the DLRR model, was not significantly different from the DLRR model despite a slight increase in the AUC. This finding is consistent with the results of Ma et al^[10]. The reason for this outcome may be that for early-stage HCC, the information provided by clinical and imaging data is limited due to the early stage and small tumour size.

However, our study has several limitations. Firstly, the retrospective nature of the study design may have inevitably introduced selection bias. Secondly, the relatively long study period may introduce bias related to treatment and imaging techniques. Nonetheless, the proposed model showed good prognostic performance in the training and test sets, suggesting that multiphase and multiregional CECT images had strong predictive value for the outcomes of ablation for HCC. Thirdly, the sample size of this study was relatively small, partly because the subjects of this study were early-stage HCC.

The DLRR model established in this study can noninvasively, efficiently, and conveniently predict ER after curative ablation in HCC patients via multiphase and multiregional CECT images. It can also stratify patients into risk subgroups based on PFS and OS rates. The successful establishment of this model provides new evidence for personalized treatment and offers close follow-up and additional treatment options for high-risk patients, thereby improving patient prognosis.

Ethics statement

This study was approved by approved by the Ethics Management Committee of First Affiliated Hospital of the University of Science and Technology of China(2021-RE-043).

Consent

The institutional ethics review board has approved our study, and the requirement for informed consent was waived because of the retrospective nature of the study.

Sources of funding

This research was Supported by Anhui Provincial Key Research and Development Plan, No. 202104j07020048.

Conflicts of interest disclosure

The authors declare that they have no financial conflicts of interest with regard to the content of this study.

Acknowledgments

The authors would like to thank Professor Jingwei Wei from the Institute of Automation, Chinese Academy of Sciences, for his support in writing this article. The authors also extend their gratitude to Professor Song Wu from Anhui University of Chinese Medicine for his guidance on statistical methodology in this manuscript.

Data availability statement

The datasets generated and/or analysed during the current study are not publicly available due to patient privacy and copyright issues but are available from the corresponding author upon reasonable request.

Country/territory of origin: China

Llovet, J. M. et al (2021)Hepatocellular carcinoma. Nature reviews 7:6
Villanueva, A. et al (2019)Hepatocellular Carcinoma. The New England journal of medicine 380:1450–1462
European Association for the Study of the Liver. Electronic address: [email protected], & European Association for the Study of the Liver(2018) EASL Clinical Practice Guidelines: Management of hepatocellular carcinoma. Journal of hepatology 69:182–236
Doyle, A. et al(2019)Outcomes of radiofrequency ablation as first-line therapy for hepatocellular carcinoma less than 3 cm in potentially transplantable patients. Journal of hepatology 70:866–873
Wang, Z. et al(2022) Microwave ablation versus laparoscopic resection as first-line therapy for solitary 3-5-cm HCC. Hepatology 76:66–77
Takayama, T. et al(2021)Surgery versus Radiofrequency Ablation for Small Hepatocellular Carcinoma: A Randomized Controlled Trial (SURF Trial). Liver cancer 11:209–218
Bosi, C. et al(2023)Understanding the causes of recurrent HCC after liver resection and radiofrequency ablation. Expert review of anticancer therapy 23:503–515
Nevola, R. et al(2023) Predictors of early and late hepatocellular carcinoma recurrence. World J Gastroenterol 29:1243–1260
Yao, L. Q. et al(2022) Correction to: Clinical Features of Recurrence After Hepatic Resection for Early-Stage Hepatocellular Carcinoma and Long-Term Survival Outcomes of Patients with Recurrence: A Multi-institutional Analysis. Ann Surg Oncol 29:5206
Ma, Q. P. et al(2021) Dynamic Contrast-Enhanced Ultrasound Radiomics for Hepatocellular Carcinoma Recurrence Prediction After Thermal Ablation. Mol Imaging Biol 23:572-585
Hu, C. et al(2021) Preoperative Gadoxetic Acid-Enhanced MRI Based Nomogram Improves Prediction of Early HCC Recurrence After Ablation Therapy. Front Oncol 11:649682
Lambin, P. et al(2012) Radiomics: extracting more information from medical images using advanced feature analysis. Eur J Cancer48:441-6
Ji, G. W. et al(2020)Radiomic Features at Contrast-enhanced CT Predict Recurrence in Early Stage Hepatocellular Carcinoma: A Multi-Institutional Study. Radiology 294:568-579
Lee, I. C. et al(2021)Evolutionary Learning-Derived Clinical-Radiomic Models for Predicting Early Recurrence of Hepatocellular Carcinoma after Resection. Liver Cancer 10:572-582
Gao, W. et al(2022)A predictive model integrating deep and radiomics features based on gadobenate dimeglumine-enhanced MRI for postoperative early recurrence of hepatocellular carcinoma. Radiol Med 127:259-271
Wang, L. et al(2023) Data-Driven Assisted Decision Making for Surgical Procedure of Hepatocellular Carcinoma Resection and Prognostic Prediction: Development and Validation of Machine Learning Models. Cancers (Basel) 15:1784
Qian, G. X. et al(2024)Computed tomography-based radiomics to predict early recurrence of hepatocellular carcinoma post-hepatectomy in patients background on cirrhosis. World J Gastroenterol 30:2128-2142
Beleù, A.et al(2022)Radiofrequency ablation of hepatocellular carcinoma: CT texture analysis of the ablated area to predict local recurrence. Eur J Radiol 150:110250
Li, Z. et al(2017) Deep Learning based Radiomics (DLR) and its usage in noninvasive IDH1 prediction for low grade glioma. Sci Rep 7:5467
Wang, F. et al(2023) A novel multimodal deep learning model for preoperative prediction of microvascular invasion and outcome in hepatocellular carcinoma. Eur J Surg Oncol 49:156-164
Wei, Z. et al(2024) Development and validation of a CT-based deep learning radiomics nomogram to predict muscle invasion in bladder cancer. Heliyon 10:e24878
Heimbach, J. K. et al(2018) AASLD guidelines for the treatment of hepatocellular carcinoma. Hepatology 67:358-380
Yang, Y. et al(2021) Early recurrence after radiofrequency ablation for hepatocellular carcinoma: a multicenter retrospective study on definition, patterns and risk factors. Int J Hyperthermia 38:437-446
Wu, J. P. et al(2022) Radiomics analysis of ultrasound to predict recurrence of hepatocellular carcinoma after microwave ablation. Int J Hyperthermia 39:595-604
Yang, X. et al(2022)Predicting hepatocellular carcinoma early recurrence after ablation based on magnetic resonance imaging radiomics nomogram. Medicine (Baltimore) 101:e32584
Yao, L. Q. et al(2022) Correction to: Clinical Features of Recurrence After Hepatic Resection for Early-Stage Hepatocellular Carcinoma and Long-Term Survival Outcomes of Patients with Recurrence: A Multi-institutional Analysis. Ann Surg Oncol 29:5206
Bosi, C. et al(2023)Understanding the causes of recurrent HCC after liver resection and radiofrequency ablation. Expert Rev Anticancer Ther 23:503-515
Yuan, C. et al(2019) Prediction early recurrence of hepatocellular carcinoma eligible for curative ablation using a Radiomics nomogram. Cancer Imaging 19:21
Shan, Q. Y. et al(2019) CT-based peritumoral radiomics signatures to predict early recurrence in hepatocellular carcinoma after curative tumor resection or ablation. Cancer Imaging 19:11
Zhou, Z. et al(2024) Prediction of preoperative microvascular invasion by dynamic radiomic analysis based on contrast-enhanced computed tomography. Abdom Radiol (NY) 49:611-624
Shi, Z. X. et al(2024)Computed tomography radiomic features and clinical factors predicting the response to first transarterial chemoembolization in intermediate-stage hepatocellular carcinoma. Hepatobiliary Pancreat Dis Int23:361-369
Chen, C. et al(2023)Multiparametric MRI-based model for prediction of local progression of hepatocellular carcinoma after thermal ablation. Cancer Med 12:17529-17540
Poplin, R. et al(2018) Prediction of cardiovascular risk factors from retinal fundus photographs via deep learning. Nat Biomed Eng 2:158-164
Zhu, Y. et al(2019) A deep learning radiomics model for preoperative grading in meningioma. Eur J Radiol 116:128-134
Liu, Z. et al(2021) Handcrafted and Deep Learning-Based Radiomic Models Can Distinguish GBM from Brain Metastasis. J Oncol 2021:5518717
Wu, J. P. et al(2022) Radiomics analysis of ultrasound to predict recurrence of hepatocellular carcinoma after microwave ablation. Int J Hyperthermia 39:595-604
Ma, Y. et al(2023) An integrated model combined intra- and peritumoral regions for predicting chemoradiation response of non small cell lung cancers based on radiomics and deep learning. Cancer Radiother 27:705-711
Zhang, J. et al(2024) Meningioma consistency assessment based on the fusion of deep learning features and radiomics features. Eur J Radiol 170:111250

Tables 1 to 4 are available in the Supplementary Files section

No competing interests reported.

Download PDF

Version 1

posted

You are reading this latest preprint version

An integrated model combined conventional radiomics and deep learning features to predict early recurrence of hepatocellular carcinoma eligible for curative ablation: a multicenter cohort study

Status:

Version 1

Abstract

Figures

INTRODUCTION

PATIENTS AND METHODS

Patient selection

Clinical data collection

Image information collection

Image segmentation

Feature extraction via radiomics and deep learning

Radiomic features and deep learning feature screening

Model building and comparison

Model validation and clinical application

Follow-up

Statistical analyses

RESULTS

Baseline characteristics of the study cohorts

Results of radiomics and deep learning feature selection

Results of model construction and comparison

Model validation and clinical application

Risk stratification

DISCUSSION

CONCLUSION

Declarations

References

Tables

Additional Declarations

Supplementary Files

Status:

Version 1